What Is Marketing Mix Modelling?

Marketing mix modelling estimates how much of your sales each marketing channel produced, using weekly spend and sales history rather than tracking individual people. It answers the budget questions that click-based attribution cannot, such as how to divide next year's money and what happens if Meta spend doubles.

The method is decades old. It is back in use because tracking people across devices got harder, and because Google and Meta now give the software away.

How it works

The model takes your weekly sales or revenue and lines it up against everything that could have moved it:

• Spend or impressions for each channel, by week.
• Things that move sales regardless of marketing, such as seasonality, public holidays, price changes and promotions.
• Outside conditions, such as inflation or unemployment.

It then separates a baseline, the sales you would have made anyway, from the sales it credits to marketing. Two ideas do most of the work. Carry-over means an ad keeps working for weeks after it runs, and saturation means each extra dollar in a channel adds less than the one before.

You get an estimated return for each channel, shown as a range, plus a curve showing where more spend stops paying.

Why it does not need cookies

The model reads totals rather than people. A week of Meta spend and a week of sales contain no personal data and no user-level tracking, so iOS privacy changes and cookie banners do not touch it.

This also means it can measure channels that clicks miss, such as radio, outdoor, sponsorship and PR.

The free tools

Two open-source packages do the heavy lifting now. Meridian comes from Google and is free to use, with a structure designed for regional data. Robyn comes from Meta's marketing science team, which describes it as experimental.

Both were built by companies that sell advertising. That does not make them wrong, but it is a good reason to test their conclusions against your own experiments.

What you need before you start

• History. Robyn's guide asks for at least two years of weekly data, or four to five years if you only have monthly numbers.
• Variation in spend. If every channel has run at the same level for two years, the model has nothing to compare. A business that has moved money around learns far more.
• Clean revenue by week. For a long sales cycle, use qualified enquiries or opportunities as the outcome, and allow for the delay to revenue.
• Several channels. With one or two channels, a simple test tells you more.
• Someone who can read the output. These are statistical models, and a wrong setting can produce a confident wrong answer.

What it cannot do

It will not tell you which ad, audience or keyword worked. Weekly totals are too broad for that, so day-to-day decisions stay in the ad platforms.

It also sees history rather than cause. Paid search spend often rises in the same weeks demand rises, and a model can credit search for sales that a TV campaign or the season created. That is why good models are calibrated with experiments: run a holdout in a few regions, then feed the measured result back in as a check.

Where it sits

Each measurement tool answers a different question:

• Platform attribution guides daily decisions on bids, ads and audiences.
• A holdout test proves whether one channel adds sales.
• Marketing mix modelling guides the annual split across all of them.

Run the model once or twice a year, in time for budget setting.

Whether it is worth doing

It earns its cost when real budget decisions depend on it: several channels, a meaningful spend and a leadership team that will actually move money based on the result. If the budget will not change whatever the model says, skip it and run a holdout instead.

Treat the output as a map of where to test next. The answer you trust is the one a test confirms.